[Bloat] Making nftables flowtable offload (SW & HW) preserve skb->priority for SQM/CAKE

[Bloat] Making nftables flowtable offload (SW & HW) preserve skb->priority for SQM/CAKE

[Julius Bairaktaris]

Hi everyone,

I have a patch series on netfilter-devel that fixes a long-standing
issue where nftables flowtable offloading breaks packet prioritization,
and it just got reviewed and tested by NVIDIA on ConnectX-7:

https://lore.kernel.org/netfilter-devel/20261004171636.3545085-1-julius@bairaktaris.de/

Right now, as soon as a connection enters a flowtable (`flow add`),
nf_flow_table drops skb->priority. That means `meta priority set` (or
IPv4 TOS-derived priority) only works for the first few slow-path
packets, after which offloaded flows lose their priority and land in the
wrong qdisc tin/band.

This series fixes that by saving skb->priority in the flowtable entry so
packets on the software fastpath keep their priority for CAKE/fq_codel,
and by emitting FLOW_ACTION_PRIORITY to hardware offload drivers
(wired up for MediaTek/Airoha PPE, plus the mlx5e handler NVIDIA just
verified so ConnectX NICs keep flows offloaded in hardware).

Thought this might be useful for anyone here running nftables flowtables
alongside SQM/QoS. Testing and feedback are very welcome!

Best,
Julius